时-空特征驱动的多轮次重构图卷积网络故障诊断方法

Fault diagnosis method driven by spatial-temporal features-based multi-round reconstructed GCN

中文关键词  故障诊断  时空特征  多轮次图重构  图卷积网络

英文关键词  fault diagnosis  spatial-temporal features  multi-round graph reconstruction  graph convolutional network

中文摘要

      近年来, 图神经网络被广泛应用于处理具有非欧结构的工业过程数据. 然而由于设备运行的过程数据常常 受到噪声和冗余信息的干扰, 如果直接使用原始信号会导致构建的图模型不够精细和准确, 从而影响后续的模型诊断性能. 针对这一问题, 本文提出了一种时–空特征驱动的多轮次重构图卷积网络(STMR-GCN)故障诊断方法. 该方法首先利用多尺度卷积神经网络与GCN对故障信号进行特征提取. 然后根据样本之间的余弦相似性对图结构进行多次重构, 重构后的图模型能够更精确地反映样本之间的连边关系, 并将得到的图模型输入到GCN进行故障种类的识别. 最后, 在东南大学(SEU)仿真数据集和真实的磨煤机数据集上进行实验, 实验结果表明所提方法与其他对比方法相比诊断精度均有提高, 从而证明STMR-GCN模型在故障诊断方面的有效性和实用性.

英文摘要

      Recently, graph neural networks have been widely used to handle industrial process data with non-Euclidean structures. However, since the process data of equipment operation is often disturbed by noises and redundant information, the direct use of raw signal to construct a graph model will result in a less accurate graph structure, thus affecting the subsequent performance of the model diagnosis. A multi-round reconstructed graph convolutional network (GCN) fault diagnosis method that driven by spatial-temporal features (STMR-GCN) is proposed. The method firstly uses multi-scale convolutional neural network with GCN for feature extraction of fault signals. Then, graph structure will be reconstructed several times according to the cosine similarity between samples, and reconstructed graph model can reflect the connected edge relationship between samples more accurately. The obtained graph model is input to GCN to realize identification of fault types. Finally, experiments are conducted on Southeast University (SEU) simulation dataset and the real coal mill dataset, and experimental results show that the proposed method improves diagnosis accuracy compared with other comparative methods, which indicates the effectiveness and feasibility of STMR-GCN model in fault diagnosis.